I built a local coding agent that learns from its wins, not just its mistakes
Most agents handle memory in one of two ways. Either they forget everything between sessions, or they "learn" by fine-tuning on a pile of past conversations and hoping the gradient sorts it out. I wanted something narrower and more honest for joe, the local-first agent shell I have been building: learn from the sessions that actually worked, and turn each one into a reusable skill I can read, edit, or delete.
This post is about the feature I just shipped to do that, and why I think the design matters more than the feature itself.
What joe is, quickly
joe is a terminal coding agent, in the spirit of Claude Code, except every model runs on my own GPU through ollama and every byte of state lives in ~/.joe-agent/ on my machine. It has the usual tools (read, write, edit, shell, grep, web), a planner, and a separate coder model it delegates to. Nothing leaves the laptop.
